mythoscode
Bahasa Indonesia
Menu

Published

One gateway, on purpose

Most coding agents let you pick a provider. MythosCode does not, and that is a design choice we enforce in code.

What the rule says

Every model call goes to the JembatanAI gateway. A second provider, a configurable base URL, a vendor login flow or a per-provider proxy counts as a regression, not a feature request.

Why

One backend means one place to account for a call. Every key, every token and every cost passes through the same door, so a bill or an incident has one place to be read from. It also means we can say what the agent talks to, and the answer is short.

How it is enforced

  • The gateway host is a build-time constant. A development override exists and is compiled out of release builds, and it only accepts loopback addresses.
  • The key comes from one environment variable. It never enters a file, a log, a fixture or a commit.
  • A static scan lists the vendor code still left to strip. Its count may only go down.
  • Tests run against an offline mock gateway with mock keys. A test that reaches the live gateway is treated as a bug.

This is the recorded output of the scan, copied from the verification record of proof M0-GW-06 in the repository’s feature list:

$ bun scripts/mythoscode/static-scan.ts --providers
providers: jembatanai only
static-scan: 0 hit(s) in 0 file(s)

What it costs us

Less choice. If you want a model the gateway does not serve, MythosCode is the wrong tool today. We think a short, checkable answer to “where does my code go” is worth more than a long provider list.

What it does not claim

It does not say the gateway is the best place for every workload, and it does not promise any model quality. It says where the calls go, and a test proves it.

Blog